[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7D3EVSCKKS0253F3A98YJPR":3,"skill-same:01M7D3EVSCKKS0253F3A98YJPR":73},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":39,"commitSha":40,"skillDirectory":41,"forksCount":13,"repositoryHtmlUrl":42,"repositoryHomepageUrl":6,"ownerHtmlUrl":43,"ownerBlogUrl":6,"ownerDescription":6,"updatedAt":44,"install":45,"resources":58},"01M7D3EVSCKKS0253F3A98YJPR","sglang",null,"Fast structured generation and serving for LLMs with RadixAttention prefix caching. Use for JSON\u002Fregex outputs, constrained decoding, agentic workflows with tool calls, or when you need 5× faster inference than vLLM with prefix sharing. Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn.","beyondai\u002Fsandbox","beyondai","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F2845958?v=4","01M7D24SBNM47BJGJ9J2HB13BP","01M7D24SCDC9DYE005XH6R553S",0,"MIT","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fsglang\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: sglang\ndescription: Fast structured generation and serving for LLMs with RadixAttention prefix caching. Use for JSON\u002Fregex outputs, constrained decoding, agentic workflows with tool calls, or when you need 5× faster inference than vLLM with prefix sharing. Powers 300,000+ GPUs at xAI, AMD, NVIDIA, and LinkedIn.\nversion: 1.0.0\nauthor: Orchestra Research\nlicense: MIT\ntags: [Inference Serving, SGLang, Structured Generation, RadixAttention, Prefix Caching, Constrained Decoding, Agents, JSON Output, Fast Inference, Production Scale]\ndependencies: [sglang, torch, transformers]\n---\n\n# SGLang\n\nHigh-performance serving framework for LLMs and VLMs with RadixAttention for automatic prefix caching.\n\n## When to use SGLang\n\n**Use SGLang when:**\n- Need structured outputs (JSON, regex, grammar)\n- Building agents with repeated prefixes (system prompts, tools)\n- Agentic workflows with function calling\n- Multi-turn conversations with shared context\n- Need faster JSON decoding (3× vs standard)\n\n**Use vLLM instead when:**\n- Simple text generation without structure\n- Don't need prefix caching\n- Want mature, widely-tested production system\n\n**Use TensorRT-LLM instead when:**\n- Maximum single-request latency (no batching needed)\n- NVIDIA-only deployment\n- Need FP8\u002FINT4 quantization on H100\n\n## Quick start\n\n### Installation\n\n```bash\n# pip install (recommended)\npip install \"sglang[all]\"\n\n# With FlashInfer (faster, CUDA 11.8\u002F12.1)\npip install sglang[all] flashinfer -i https:\u002F\u002Fflashinfer.ai\u002Fwhl\u002Fcu121\u002Ftorch2.4\u002F\n\n# From source\ngit clone https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang.git\ncd sglang\npip install -e \"python[all]\"\n```\n\n### Launch server\n\n```bash\n# Basic server (Llama 3-8B)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000\n\n# With RadixAttention (automatic prefix caching)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000 \\\n    --enable-radix-cache  # Default: enabled\n\n# Multi-GPU (tensor parallelism)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-70B-Instruct \\\n    --tp 4 \\\n    --port 30000\n```\n\n### Basic inference\n\n```python\nimport sglang as sgl\n\n# Set backend\nsgl.set_default_backend(sgl.OpenAI(\"http:\u002F\u002Flocalhost:30000\u002Fv1\"))\n\n# Simple generation\n@sgl.function\ndef simple_gen(s, question):\n    s += \"Q: \" + question + \"\\n\"\n    s += \"A:\" + sgl.gen(\"answer\", max_tokens=100)\n\n# Run\nstate = simple_gen.run(question=\"What is the capital of France?\")\nprint(state[\"answer\"])\n# Output: \"The capital of France is Paris.\"\n```\n\n### Structured JSON output\n\n```python\nimport sglang as sgl\n\n@sgl.function\ndef extract_person(s, text):\n    s += f\"Extract person information from: {text}\\n\"\n    s += \"Output JSON:\\n\"\n\n    # Constrained JSON generation\n    s += sgl.gen(\n        \"json_output\",\n        max_tokens=200,\n        regex=r'\\{\"name\": \"[^\"]+\", \"age\": \\d+, \"occupation\": \"[^\"]+\"\\}'\n    )\n\n# Run\nstate = extract_person.run(\n    text=\"John Smith is a 35-year-old software engineer.\"\n)\nprint(state[\"json_output\"])\n# Output: {\"name\": \"John Smith\", \"age\": 35, \"occupation\": \"software engineer\"}\n```\n\n## RadixAttention (Key Innovation)\n\n**What it does**: Automatically caches and reuses common prefixes across requests.\n\n**Performance**:\n- **5× faster** for agentic workloads with shared system prompts\n- **10× faster** for few-shot prompting with repeated examples\n- **Zero configuration** - works automatically\n\n**How it works**:\n1. Builds radix tree of all processed tokens\n2. Automatically detects shared prefixes\n3. Reuses KV cache for matching prefixes\n4. Only computes new tokens\n\n**Example** (Agent with system prompt):\n\n```\nRequest 1: [SYSTEM_PROMPT] + \"What's the weather?\"\n→ Computes full prompt (1000 tokens)\n\nRequest 2: [SAME_SYSTEM_PROMPT] + \"Book a flight\"\n→ Reuses system prompt KV cache (998 tokens)\n→ Only computes 2 new tokens\n→ 5× faster!\n```\n\n## Structured generation patterns\n\n### JSON with schema\n\n```python\n@sgl.function\ndef structured_extraction(s, article):\n    s += f\"Article: {article}\\n\\n\"\n    s += \"Extract key information as JSON:\\n\"\n\n    # JSON schema constraint\n    schema = {\n        \"type\": \"object\",\n        \"properties\": {\n            \"title\": {\"type\": \"string\"},\n            \"author\": {\"type\": \"string\"},\n            \"summary\": {\"type\": \"string\"},\n            \"sentiment\": {\"type\": \"string\", \"enum\": [\"positive\", \"negative\", \"neutral\"]}\n        },\n        \"required\": [\"title\", \"author\", \"summary\", \"sentiment\"]\n    }\n\n    s += sgl.gen(\"info\", max_tokens=300, json_schema=schema)\n\nstate = structured_extraction.run(article=\"...\")\nprint(state[\"info\"])\n# Output: Valid JSON matching schema\n```\n\n### Regex-constrained generation\n\n```python\n@sgl.function\ndef extract_email(s, text):\n    s += f\"Extract email from: {text}\\n\"\n    s += \"Email: \"\n\n    # Email regex pattern\n    s += sgl.gen(\n        \"email\",\n        max_tokens=50,\n        regex=r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}'\n    )\n\nstate = extract_email.run(text=\"Contact john.doe@example.com for details\")\nprint(state[\"email\"])\n# Output: \"john.doe@example.com\"\n```\n\n### Grammar-based generation\n\n```python\n@sgl.function\ndef generate_code(s, description):\n    s += f\"Generate Python code for: {description}\\n\"\n    s += \"```python\\n\"\n\n    # EBNF grammar for Python\n    python_grammar = \"\"\"\n    ?start: function_def\n    function_def: \"def\" NAME \"(\" [parameters] \"):\" suite\n    parameters: parameter (\",\" parameter)*\n    parameter: NAME\n    suite: simple_stmt | NEWLINE INDENT stmt+ DEDENT\n    \"\"\"\n\n    s += sgl.gen(\"code\", max_tokens=200, grammar=python_grammar)\n    s += \"\\n```\"\n```\n\n## Agent workflows with function calling\n\n```python\nimport sglang as sgl\n\n# Define tools\ntools = [\n    {\n        \"name\": \"get_weather\",\n        \"description\": \"Get weather for a location\",\n        \"parameters\": {\n            \"type\": \"object\",\n            \"properties\": {\n                \"location\": {\"type\": \"string\"}\n            }\n        }\n    },\n    {\n        \"name\": \"book_flight\",\n        \"description\": \"Book a flight\",\n        \"parameters\": {\n            \"type\": \"object\",\n            \"properties\": {\n                \"from\": {\"type\": \"string\"},\n                \"to\": {\"type\": \"string\"},\n                \"date\": {\"type\": \"string\"}\n            }\n        }\n    }\n]\n\n@sgl.function\ndef agent_workflow(s, user_query, tools):\n    # System prompt (cached with RadixAttention)\n    s += \"You are a helpful assistant with access to tools.\\n\"\n    s += f\"Available tools: {tools}\\n\\n\"\n\n    # User query\n    s += f\"User: {user_query}\\n\"\n    s += \"Assistant: \"\n\n    # Generate with function calling\n    s += sgl.gen(\n        \"response\",\n        max_tokens=200,\n        tools=tools,  # SGLang handles tool call format\n        stop=[\"User:\", \"\\n\\n\"]\n    )\n\n# Multiple queries reuse system prompt\nstate1 = agent_workflow.run(\n    user_query=\"What's the weather in NYC?\",\n    tools=tools\n)\n# First call: Computes full system prompt\n\nstate2 = agent_workflow.run(\n    user_query=\"Book a flight to LA\",\n    tools=tools\n)\n# Second call: Reuses system prompt (5× faster)\n```\n\n## Performance benchmarks\n\n### RadixAttention speedup\n\n**Few-shot prompting** (10 examples in prompt):\n- vLLM: 2.5 sec\u002Frequest\n- SGLang: **0.25 sec\u002Frequest** (10× faster)\n- Throughput: 4× higher\n\n**Agent workflows** (1000-token system prompt):\n- vLLM: 1.8 sec\u002Frequest\n- SGLang: **0.35 sec\u002Frequest** (5× faster)\n\n**JSON decoding**:\n- Standard: 45 tok\u002Fs\n- SGLang: **135 tok\u002Fs** (3× faster)\n\n### Throughput (Llama 3-8B, A100)\n\n| Workload | vLLM | SGLang | Speedup |\n|----------|------|--------|---------|\n| Simple generation | 2500 tok\u002Fs | 2800 tok\u002Fs | 1.12× |\n| Few-shot (10 examples) | 500 tok\u002Fs | 5000 tok\u002Fs | 10× |\n| Agent (tool calls) | 800 tok\u002Fs | 4000 tok\u002Fs | 5× |\n| JSON output | 600 tok\u002Fs | 2400 tok\u002Fs | 4× |\n\n## Multi-turn conversations\n\n```python\n@sgl.function\ndef multi_turn_chat(s, history, new_message):\n    # System prompt (always cached)\n    s += \"You are a helpful AI assistant.\\n\\n\"\n\n    # Conversation history (cached as it grows)\n    for msg in history:\n        s += f\"{msg['role']}: {msg['content']}\\n\"\n\n    # New user message (only new part)\n    s += f\"User: {new_message}\\n\"\n    s += \"Assistant: \"\n    s += sgl.gen(\"response\", max_tokens=200)\n\n# Turn 1\nhistory = []\nstate = multi_turn_chat.run(history=history, new_message=\"Hi there!\")\nhistory.append({\"role\": \"User\", \"content\": \"Hi there!\"})\nhistory.append({\"role\": \"Assistant\", \"content\": state[\"response\"]})\n\n# Turn 2 (reuses Turn 1 KV cache)\nstate = multi_turn_chat.run(history=history, new_message=\"What's 2+2?\")\n# Only computes new message (much faster!)\n\n# Turn 3 (reuses Turn 1 + Turn 2 KV cache)\nstate = multi_turn_chat.run(history=history, new_message=\"Tell me a joke\")\n# Progressively faster as history grows\n```\n\n## Advanced features\n\n### Speculative decoding\n\n```bash\n# Launch with draft model (2-3× faster)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-70B-Instruct \\\n    --speculative-model meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --speculative-num-steps 5\n```\n\n### Multi-modal (vision models)\n\n```python\n@sgl.function\ndef describe_image(s, image_path):\n    s += sgl.image(image_path)\n    s += \"Describe this image in detail: \"\n    s += sgl.gen(\"description\", max_tokens=200)\n\nstate = describe_image.run(image_path=\"photo.jpg\")\nprint(state[\"description\"])\n```\n\n### Batching and parallel requests\n\n```python\n# Automatic batching (continuous batching)\nstates = sgl.run_batch(\n    [\n        simple_gen.bind(question=\"What is AI?\"),\n        simple_gen.bind(question=\"What is ML?\"),\n        simple_gen.bind(question=\"What is DL?\"),\n    ]\n)\n\n# All 3 processed in single batch (efficient)\n```\n\n## OpenAI-compatible API\n\n```bash\n# Start server with OpenAI API\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000\n\n# Use with OpenAI client\ncurl http:\u002F\u002Flocalhost:30000\u002Fv1\u002Fchat\u002Fcompletions \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\n    \"model\": \"default\",\n    \"messages\": [\n      {\"role\": \"system\", \"content\": \"You are helpful\"},\n      {\"role\": \"user\", \"content\": \"Hello\"}\n    ],\n    \"temperature\": 0.7,\n    \"max_tokens\": 100\n  }'\n\n# Works with OpenAI Python SDK\nfrom openai import OpenAI\nclient = OpenAI(base_url=\"http:\u002F\u002Flocalhost:30000\u002Fv1\", api_key=\"EMPTY\")\n\nresponse = client.chat.completions.create(\n    model=\"default\",\n    messages=[{\"role\": \"user\", \"content\": \"Hello\"}]\n)\n```\n\n## Supported models\n\n**Text models**:\n- Llama 2, Llama 3, Llama 3.1, Llama 3.2\n- Mistral, Mixtral\n- Qwen, Qwen2, QwQ\n- DeepSeek-V2, DeepSeek-V3\n- Gemma, Phi-3\n\n**Vision models**:\n- LLaVA, LLaVA-OneVision\n- Phi-3-Vision\n- Qwen2-VL\n\n**100+ models** from HuggingFace\n\n## Hardware support\n\n**NVIDIA**: A100, H100, L4, T4 (CUDA 11.8+)\n**AMD**: MI300, MI250 (ROCm 6.0+)\n**Intel**: Xeon with GPU (coming soon)\n**Apple**: M1\u002FM2\u002FM3 via MPS (experimental)\n\n## References\n\n- **[Structured Generation Guide](references\u002Fstructured-generation.md)** - JSON schemas, regex, grammars, validation\n- **[RadixAttention Deep Dive](references\u002Fradix-attention.md)** - How it works, optimization, benchmarks\n- **[Production Deployment](references\u002Fdeployment.md)** - Multi-GPU, monitoring, autoscaling\n\n## Resources\n\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\n- **Docs**: https:\u002F\u002Fsgl-project.github.io\u002F\n- **Paper**: RadixAttention (arXiv:2312.07104)\n- **Discord**: https:\u002F\u002Fdiscord.gg\u002Fsglang\n\n\n","# SGLang\n\nHigh-performance serving framework for LLMs and VLMs with RadixAttention for automatic prefix caching.\n\n## When to use SGLang\n\n**Use SGLang when:**\n- Need structured outputs (JSON, regex, grammar)\n- Building agents with repeated prefixes (system prompts, tools)\n- Agentic workflows with function calling\n- Multi-turn conversations with shared context\n- Need faster JSON decoding (3× vs standard)\n\n**Use vLLM instead when:**\n- Simple text generation without structure\n- Don't need prefix caching\n- Want mature, widely-tested production system\n\n**Use TensorRT-LLM instead when:**\n- Maximum single-request latency (no batching needed)\n- NVIDIA-only deployment\n- Need FP8\u002FINT4 quantization on H100\n\n## Quick start\n\n### Installation\n\n```bash\n# pip install (recommended)\npip install \"sglang[all]\"\n\n# With FlashInfer (faster, CUDA 11.8\u002F12.1)\npip install sglang[all] flashinfer -i https:\u002F\u002Fflashinfer.ai\u002Fwhl\u002Fcu121\u002Ftorch2.4\u002F\n\n# From source\ngit clone https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang.git\ncd sglang\npip install -e \"python[all]\"\n```\n\n### Launch server\n\n```bash\n# Basic server (Llama 3-8B)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000\n\n# With RadixAttention (automatic prefix caching)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000 \\\n    --enable-radix-cache  # Default: enabled\n\n# Multi-GPU (tensor parallelism)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-70B-Instruct \\\n    --tp 4 \\\n    --port 30000\n```\n\n### Basic inference\n\n```python\nimport sglang as sgl\n\n# Set backend\nsgl.set_default_backend(sgl.OpenAI(\"http:\u002F\u002Flocalhost:30000\u002Fv1\"))\n\n# Simple generation\n@sgl.function\ndef simple_gen(s, question):\n    s += \"Q: \" + question + \"\\n\"\n    s += \"A:\" + sgl.gen(\"answer\", max_tokens=100)\n\n# Run\nstate = simple_gen.run(question=\"What is the capital of France?\")\nprint(state[\"answer\"])\n# Output: \"The capital of France is Paris.\"\n```\n\n### Structured JSON output\n\n```python\nimport sglang as sgl\n\n@sgl.function\ndef extract_person(s, text):\n    s += f\"Extract person information from: {text}\\n\"\n    s += \"Output JSON:\\n\"\n\n    # Constrained JSON generation\n    s += sgl.gen(\n        \"json_output\",\n        max_tokens=200,\n        regex=r'\\{\"name\": \"[^\"]+\", \"age\": \\d+, \"occupation\": \"[^\"]+\"\\}'\n    )\n\n# Run\nstate = extract_person.run(\n    text=\"John Smith is a 35-year-old software engineer.\"\n)\nprint(state[\"json_output\"])\n# Output: {\"name\": \"John Smith\", \"age\": 35, \"occupation\": \"software engineer\"}\n```\n\n## RadixAttention (Key Innovation)\n\n**What it does**: Automatically caches and reuses common prefixes across requests.\n\n**Performance**:\n- **5× faster** for agentic workloads with shared system prompts\n- **10× faster** for few-shot prompting with repeated examples\n- **Zero configuration** - works automatically\n\n**How it works**:\n1. Builds radix tree of all processed tokens\n2. Automatically detects shared prefixes\n3. Reuses KV cache for matching prefixes\n4. Only computes new tokens\n\n**Example** (Agent with system prompt):\n\n```\nRequest 1: [SYSTEM_PROMPT] + \"What's the weather?\"\n→ Computes full prompt (1000 tokens)\n\nRequest 2: [SAME_SYSTEM_PROMPT] + \"Book a flight\"\n→ Reuses system prompt KV cache (998 tokens)\n→ Only computes 2 new tokens\n→ 5× faster!\n```\n\n## Structured generation patterns\n\n### JSON with schema\n\n```python\n@sgl.function\ndef structured_extraction(s, article):\n    s += f\"Article: {article}\\n\\n\"\n    s += \"Extract key information as JSON:\\n\"\n\n    # JSON schema constraint\n    schema = {\n        \"type\": \"object\",\n        \"properties\": {\n            \"title\": {\"type\": \"string\"},\n            \"author\": {\"type\": \"string\"},\n            \"summary\": {\"type\": \"string\"},\n            \"sentiment\": {\"type\": \"string\", \"enum\": [\"positive\", \"negative\", \"neutral\"]}\n        },\n        \"required\": [\"title\", \"author\", \"summary\", \"sentiment\"]\n    }\n\n    s += sgl.gen(\"info\", max_tokens=300, json_schema=schema)\n\nstate = structured_extraction.run(article=\"...\")\nprint(state[\"info\"])\n# Output: Valid JSON matching schema\n```\n\n### Regex-constrained generation\n\n```python\n@sgl.function\ndef extract_email(s, text):\n    s += f\"Extract email from: {text}\\n\"\n    s += \"Email: \"\n\n    # Email regex pattern\n    s += sgl.gen(\n        \"email\",\n        max_tokens=50,\n        regex=r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\\.[a-zA-Z]{2,}'\n    )\n\nstate = extract_email.run(text=\"Contact john.doe@example.com for details\")\nprint(state[\"email\"])\n# Output: \"john.doe@example.com\"\n```\n\n### Grammar-based generation\n\n```python\n@sgl.function\ndef generate_code(s, description):\n    s += f\"Generate Python code for: {description}\\n\"\n    s += \"```python\\n\"\n\n    # EBNF grammar for Python\n    python_grammar = \"\"\"\n    ?start: function_def\n    function_def: \"def\" NAME \"(\" [parameters] \"):\" suite\n    parameters: parameter (\",\" parameter)*\n    parameter: NAME\n    suite: simple_stmt | NEWLINE INDENT stmt+ DEDENT\n    \"\"\"\n\n    s += sgl.gen(\"code\", max_tokens=200, grammar=python_grammar)\n    s += \"\\n```\"\n```\n\n## Agent workflows with function calling\n\n```python\nimport sglang as sgl\n\n# Define tools\ntools = [\n    {\n        \"name\": \"get_weather\",\n        \"description\": \"Get weather for a location\",\n        \"parameters\": {\n            \"type\": \"object\",\n            \"properties\": {\n                \"location\": {\"type\": \"string\"}\n            }\n        }\n    },\n    {\n        \"name\": \"book_flight\",\n        \"description\": \"Book a flight\",\n        \"parameters\": {\n            \"type\": \"object\",\n            \"properties\": {\n                \"from\": {\"type\": \"string\"},\n                \"to\": {\"type\": \"string\"},\n                \"date\": {\"type\": \"string\"}\n            }\n        }\n    }\n]\n\n@sgl.function\ndef agent_workflow(s, user_query, tools):\n    # System prompt (cached with RadixAttention)\n    s += \"You are a helpful assistant with access to tools.\\n\"\n    s += f\"Available tools: {tools}\\n\\n\"\n\n    # User query\n    s += f\"User: {user_query}\\n\"\n    s += \"Assistant: \"\n\n    # Generate with function calling\n    s += sgl.gen(\n        \"response\",\n        max_tokens=200,\n        tools=tools,  # SGLang handles tool call format\n        stop=[\"User:\", \"\\n\\n\"]\n    )\n\n# Multiple queries reuse system prompt\nstate1 = agent_workflow.run(\n    user_query=\"What's the weather in NYC?\",\n    tools=tools\n)\n# First call: Computes full system prompt\n\nstate2 = agent_workflow.run(\n    user_query=\"Book a flight to LA\",\n    tools=tools\n)\n# Second call: Reuses system prompt (5× faster)\n```\n\n## Performance benchmarks\n\n### RadixAttention speedup\n\n**Few-shot prompting** (10 examples in prompt):\n- vLLM: 2.5 sec\u002Frequest\n- SGLang: **0.25 sec\u002Frequest** (10× faster)\n- Throughput: 4× higher\n\n**Agent workflows** (1000-token system prompt):\n- vLLM: 1.8 sec\u002Frequest\n- SGLang: **0.35 sec\u002Frequest** (5× faster)\n\n**JSON decoding**:\n- Standard: 45 tok\u002Fs\n- SGLang: **135 tok\u002Fs** (3× faster)\n\n### Throughput (Llama 3-8B, A100)\n\n| Workload | vLLM | SGLang | Speedup |\n|----------|------|--------|---------|\n| Simple generation | 2500 tok\u002Fs | 2800 tok\u002Fs | 1.12× |\n| Few-shot (10 examples) | 500 tok\u002Fs | 5000 tok\u002Fs | 10× |\n| Agent (tool calls) | 800 tok\u002Fs | 4000 tok\u002Fs | 5× |\n| JSON output | 600 tok\u002Fs | 2400 tok\u002Fs | 4× |\n\n## Multi-turn conversations\n\n```python\n@sgl.function\ndef multi_turn_chat(s, history, new_message):\n    # System prompt (always cached)\n    s += \"You are a helpful AI assistant.\\n\\n\"\n\n    # Conversation history (cached as it grows)\n    for msg in history:\n        s += f\"{msg['role']}: {msg['content']}\\n\"\n\n    # New user message (only new part)\n    s += f\"User: {new_message}\\n\"\n    s += \"Assistant: \"\n    s += sgl.gen(\"response\", max_tokens=200)\n\n# Turn 1\nhistory = []\nstate = multi_turn_chat.run(history=history, new_message=\"Hi there!\")\nhistory.append({\"role\": \"User\", \"content\": \"Hi there!\"})\nhistory.append({\"role\": \"Assistant\", \"content\": state[\"response\"]})\n\n# Turn 2 (reuses Turn 1 KV cache)\nstate = multi_turn_chat.run(history=history, new_message=\"What's 2+2?\")\n# Only computes new message (much faster!)\n\n# Turn 3 (reuses Turn 1 + Turn 2 KV cache)\nstate = multi_turn_chat.run(history=history, new_message=\"Tell me a joke\")\n# Progressively faster as history grows\n```\n\n## Advanced features\n\n### Speculative decoding\n\n```bash\n# Launch with draft model (2-3× faster)\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-70B-Instruct \\\n    --speculative-model meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --speculative-num-steps 5\n```\n\n### Multi-modal (vision models)\n\n```python\n@sgl.function\ndef describe_image(s, image_path):\n    s += sgl.image(image_path)\n    s += \"Describe this image in detail: \"\n    s += sgl.gen(\"description\", max_tokens=200)\n\nstate = describe_image.run(image_path=\"photo.jpg\")\nprint(state[\"description\"])\n```\n\n### Batching and parallel requests\n\n```python\n# Automatic batching (continuous batching)\nstates = sgl.run_batch(\n    [\n        simple_gen.bind(question=\"What is AI?\"),\n        simple_gen.bind(question=\"What is ML?\"),\n        simple_gen.bind(question=\"What is DL?\"),\n    ]\n)\n\n# All 3 processed in single batch (efficient)\n```\n\n## OpenAI-compatible API\n\n```bash\n# Start server with OpenAI API\npython -m sglang.launch_server \\\n    --model-path meta-llama\u002FMeta-Llama-3-8B-Instruct \\\n    --port 30000\n\n# Use with OpenAI client\ncurl http:\u002F\u002Flocalhost:30000\u002Fv1\u002Fchat\u002Fcompletions \\\n  -H \"Content-Type: application\u002Fjson\" \\\n  -d '{\n    \"model\": \"default\",\n    \"messages\": [\n      {\"role\": \"system\", \"content\": \"You are helpful\"},\n      {\"role\": \"user\", \"content\": \"Hello\"}\n    ],\n    \"temperature\": 0.7,\n    \"max_tokens\": 100\n  }'\n\n# Works with OpenAI Python SDK\nfrom openai import OpenAI\nclient = OpenAI(base_url=\"http:\u002F\u002Flocalhost:30000\u002Fv1\", api_key=\"EMPTY\")\n\nresponse = client.chat.completions.create(\n    model=\"default\",\n    messages=[{\"role\": \"user\", \"content\": \"Hello\"}]\n)\n```\n\n## Supported models\n\n**Text models**:\n- Llama 2, Llama 3, Llama 3.1, Llama 3.2\n- Mistral, Mixtral\n- Qwen, Qwen2, QwQ\n- DeepSeek-V2, DeepSeek-V3\n- Gemma, Phi-3\n\n**Vision models**:\n- LLaVA, LLaVA-OneVision\n- Phi-3-Vision\n- Qwen2-VL\n\n**100+ models** from HuggingFace\n\n## Hardware support\n\n**NVIDIA**: A100, H100, L4, T4 (CUDA 11.8+)\n**AMD**: MI300, MI250 (ROCm 6.0+)\n**Intel**: Xeon with GPU (coming soon)\n**Apple**: M1\u002FM2\u002FM3 via MPS (experimental)\n\n## References\n\n- **[Structured Generation Guide](references\u002Fstructured-generation.md)** - JSON schemas, regex, grammars, validation\n- **[RadixAttention Deep Dive](references\u002Fradix-attention.md)** - How it works, optimization, benchmarks\n- **[Production Deployment](references\u002Fdeployment.md)** - Multi-GPU, monitoring, autoscaling\n\n## Resources\n\n- **GitHub**: https:\u002F\u002Fgithub.com\u002Fsgl-project\u002Fsglang\n- **Docs**: https:\u002F\u002Fsgl-project.github.io\u002F\n- **Paper**: RadixAttention (arXiv:2312.07104)\n- **Discord**: https:\u002F\u002Fdiscord.gg\u002Fsglang\n\n\n",{"name":5,"tags":23,"author":34,"license":14,"version":35,"description":7,"dependencies":36},[24,25,26,27,28,29,30,31,32,33],"Inference Serving","SGLang","Structured Generation","RadixAttention","Prefix Caching","Constrained Decoding","Agents","JSON Output","Fast Inference","Production Scale","Orchestra Research","1.0.0",[5,37,38],"torch","transformers","main","46e03791269608c5ac886a64c08f5d94147e8786","labs\u002F.opencode\u002Fskills\u002Fsglang","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox","https:\u002F\u002Fgithub.com\u002Fbeyondai","2026-11-08 04:18:42",{"source":8,"skillName":5,"commands":46,"prompt":56,"zipAvailable":57},[47,50,53],{"tool":48,"command":49},"npx","npx skills add beyondai\u002Fsandbox --skill sglang",{"tool":51,"command":52},"bunx","bunx skills add beyondai\u002Fsandbox --skill sglang",{"tool":54,"command":55},"pnpm","pnpm dlx skills add beyondai\u002Fsandbox --skill sglang","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fsglang\u002FSKILL.md",true,[59,65,69],{"path":60,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":63,"sourceHtmlUrl":64},"references\u002Fdeployment.md","REFERENCE","参考文档",10020,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fsglang\u002Freferences\u002Fdeployment.md",{"path":66,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":67,"sourceHtmlUrl":68},"references\u002Fradix-attention.md",11005,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fsglang\u002Freferences\u002Fradix-attention.md",{"path":70,"resourceType":61,"resourceTypeDesc":62,"sizeBytes":71,"sourceHtmlUrl":72},"references\u002Fstructured-generation.md",13468,"https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fsglang\u002Freferences\u002Fstructured-generation.md",[74,81,88,95,102,109],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":18},"01M7D3GMPGJZRG8EPY404YF8TR","weights-and-biases","Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps platform","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fweights-and-biases\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":18},"01M7D3GFZH440V2J532BSZMPVP","serving-llms-vllm","Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. Use when deploying production LLM APIs, optimizing inference latency\u002Fthroughput, or serving models with limited GPU memory. Supports OpenAI-compatible endpoints, quantization (GPTQ\u002FAWQ\u002FFP8), and tensor parallelism.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fvllm\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":18},"01M7D3GAE8MQCPGF72EP9G9GKY","verl-rl-training","Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). Use when implementing RLHF, GRPO, PPO, or other RL algorithms for LLM post-training at scale with flexible infrastructure backends.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Fverl\u002FSKILL.md",[],[],{"skillNo":96,"name":97,"nameZh":6,"description":98,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":18},"01M7D3G7KKNE5RGNYJNXV4F0MM","unsloth","Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA\u002FQLoRA optimization","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Funsloth\u002FSKILL.md",[],[],{"skillNo":103,"name":104,"nameZh":6,"description":105,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":106,"categories":107,"keywords":108,"publishedAt":18},"01M7D3G2RCZQ08GS0KAW5CBWSC","fine-tuning-with-trl","Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO\u002FGRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Ftrl-fine-tuning\u002FSKILL.md",[],[],{"skillNo":110,"name":111,"nameZh":6,"description":112,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":6,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":113,"categories":114,"keywords":115,"publishedAt":18},"01M7D3FX08BPB172CYGM597NZF","torchforge-rl-training","Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. Use when you want clean RL abstractions, easy algorithm experimentation, or scalable training with Monarch and TorchTitan.","https:\u002F\u002Fgithub.com\u002Fbeyondai\u002Fsandbox\u002Fblob\u002Fmain\u002Flabs\u002F.opencode\u002Fskills\u002Ftorchforge\u002FSKILL.md",[],[]]